Search

In this paper, we propose a feature affinity (FA) assisted knowledge distillation (KD) method to improve quantization-aware training of deep neural networks (DNN). The FA loss on intermediate feature maps of DNNs plays the role of teaching middle steps of a solution to a student instead of only giving final answers in the conventional KD where the loss acts on the network logits at the output level. Combining logit loss and FA loss, we found via convolutional network experiments on CIFAR-10/100, and Tiny ImageNet data sets that the quantized student network receives stronger supervision than from the labeled ground-truth data. The resulting FA quantization-distillation (FAQD), trained to convergence with a cosine annealing scheduler for 200 epochs, is capable of compressing models on label-free data up to or exceeding the accuracy levels of their full precision counterparts, which brings immediate practical benefits as pre-trained teacher models are readily available and unlabeled data are abundant. In contrast, data labeling is often laborious and expensive. Finally, we propose and prove error estimates for a fast feature affinity (FFA) loss function that accurately approximates FA loss at a lower order of computational complexity, which helps speed up training for high resolution image input. Source codes are available at: https://github.com/lzj994/FAQD

This paper proposed an input-parallel and output-series (IPOS) LLC resonant converter for applications with high voltage gain and low ripple, which considers the sharing of submodule output voltages. The resonant converter proposed has an integrated magnetics, which improves the utilization of the core. The reason is because magnetic fluxes can be superimposed or cancelled each other when the resonant currents are interleaved. Moreover, the influence of parameter mismatch in the resonant tank on voltage and current sharing in IPOS-LLC was analyzed based on the passive impedance theory. The results show that, by selecting smaller quality factors and using N-connections in high-voltage applications, the sharing of output voltage among modules can be well reflected, despite tolerances of the resonance parameters. Meanwhile, the voltage stress distribution of the rectifier diodes in each submodule is optimized, which improves the reliability of the circuit. By using the aforementioned conclusions, a prototype with 90 kHz, 40 V/4.8 kV, and 1.15 kW is established. The measured output voltage mismatch of 3.1% and voltage ripple rate of 0.16% at 3.7 nF output capacitance proves the effectiveness and correctness of the proposed resonant converter design.

The purpose of this study was to assess the impact of Artificial Intelligence (AI) on education. Premised on a narrative and framework for assessing AI identified from a preliminary analysis, the scope of the study was limited to the application and effects of AI in administration, instruction, and learning. A qualitative research approach, leveraging the use of literature review as a research design and approach was used and effectively facilitated the realization of the study purpose. Artificial intelligence is a field of study and the resulting innovations and developments that have culminated in computers, machines, and other artifacts having human-like intelligence characterized by cognitive abilities, learning, adaptability, and decision-making capabilities. The study ascertained that AI has extensively been adopted and used in education, particularly by education institutions, in different forms. AI initially took the form of computer and computer related technologies, transitioning to web-based and online intelligent education systems, and ultimately with the use of embedded computer systems, together with other technologies, the use of humanoid robots and web-based chatbots to perform instructors' duties and functions independently or with instructors. Using these platforms, instructors have been able to perform different administrative functions, such as reviewing and grading students' assignments more effectively and efficiently, and achieve higher quality in their teaching activities. On the other hand, because the systems leverage machine learning and adaptability, curriculum and content has been customized and personalized in line with students' needs, which has fostered uptake and retention, thereby improving learners experience and overall quality of learning.

Device-to-Device (D2D) communication with caching technology has emerged as a promising technique for offloading the traffic and boosting the throughput of the fifth generation (5G) cellular networks. The combined impact of cache memory size of user equipments (UEs) and content sizes, which are two crucial factors in D2D-assisted caching networks, were usually ignored in the existing researches. In this paper, an optimization algorithm is proposed to maximize the cache hit probability and cache-aided throughput, with the consideration of various cache memory sizes of UEs and content sizes. Firstly, users are grouped according to the cache memory sizes and the content sizes. Then the general mathematical expressions for the optimization of cache hit probability and cache-aided throughput with the constraints of cache memory sizes and content sizes are obtained. Subsequently, a Packet Cache Strategy (PCS) algorithm is proposed to obtain the caching probability matrix with the maximum cache hit probability and cache-aided throughout by taking user caching probability of a file as a variable. Finally, numerical results show that the sizes of the requested files affect the caching willingness of users, and the proposed PCS can achieve the highest cache hit probability and the best cache-aided throughput comparing with two other existing methods.

Doubly fed induction generator (DFIG) may suffer from huge disturbance owing to the uncertainty of generator parameters and load fluctuation during operation process. Hence, in the context of virtual synchronous generator (VSG) technology and its power-frequency control characteristics, this paper presents a VSG-based nonlinear adaptive robust controller (NARC), aiming to enhance the comprehensive regulation performance of DFIG. Firstly, a nonlinear robust adaptive controller with disturbance rejection and uncertainty adaptive learning ability is designed by establishing an improved dynamic model of the rotor motion equation in VSG, so as to achieve the optimal regulation of DFIG with robustness and adaptability. Secondly, in order to realize the fast dynamic response and autonomous synchronization ability in DFIG, a modified control strategy of the grid side converter (GSC) in DFIG is constructed by voltage and current dual-loop control (DLC) and VSG, which not only retains the significant dynamic response ability of the current inner-loop in DLC, but also improves the inertia and damping of the system. Finally, simulation results indicate that the modified control strategy exhibits satisfied dynamic performance and reliability, and reveals the robustness and adaptability of the control strategy in response to external random interferences and internal uncertain parameters.

The existing physical-layer network coding (PNC) only considers channel codes of a constant coding rate. In this paper, we propose a rate-compatible (RC) channel coding scheme for PNC in two-way relay systems. Our contributions are as follows. First, we propose a two-edge-type-low-density-paritycheck-aided PNC (JTNC) scheme due to its simple implementation and excellent performance. Second, we design an RC scheme, i.e., choice of locations and symbols, for the proposed JTNC in the relay. The simulation results show that the proposed RC-JTNC has good error correction performance and achieves large gains over the existing puncturing algorithms aided PNC over a wide range of rates.

Different storage capacities and various file sizes are two crucial factors which impact the effectiveness of cache placement in device-to-device networks. However, most of existing literature assumes all users have homogeneous cache capability and all involved file sizes are identical. In this paper, files are placed into groups according to their different sizes and the storage capacities are various among all the involved user devices. Considering these two vital factors, group caching scheme (GCS) algorithm for the optimal cache strategy is investigated by formulating an optimization problem to maximize the cache hit probability solved by using Karush-Kuhn-Tucker conditions. Furthermore, random linear network coding-based GCS is developed to eliminate the negative impact of file size to cache hit probability due to the observation that the probability of small size files being cached may be higher than that of large size but more popular files. Finally, simulation results show the effectiveness of the proposal which outperforms the existing schemes.